The Impact of Population Size, Number of Children, and Number of Reference Points on the Performance of NSGA-III

摘要

We investigate the impact of three control parameters (the population size μ, the number of children λ, and the number of reference points H) on the performance of Nondominated Sorting Genetic Algorithm III (NSGA-III). In the past few years, many efficient Multi-Objective Evolutionary Algorithms (MOEAs) for Many-Objective Optimization Problems (MaOPs) have been proposed, but their control parameters have been poorly analyzed. The recently proposed NSGA-III is one of most promising MOEAs for MaOPs. It is widely believed that NSGA-III is almost parameter-less and requires setting only one control parameter (H), and the value of μ and λ can be set to μ = λ ≈ H as described in the original NSGA-III paper. However, the experimental results in this paper show that suitable parameter settings of μ, λ, and H values differ from each other as well as their widely used parameter settings. Also, the performance of NSGA-III significantly depends on them. Thus, the usually used parameter settings of NSGA-III (i.e., μ = λ ≈ H) might be unsuitable in many cases, and μ, λ, and H require a particular parameter tuning to realize the best performance of NSGA-III.